Xiaohu AI Daily — 2026-05-25

🌟 Today's Headline

OpenJDK News Roundup: Vector API, Compact Object Headers and G1GC as Default in JDK 27

There was a flurry of activity in the OpenJDK ecosystem during the week of May 18th, 2026, highlighting three JEPs elevated from Proposed to Target to Targeted and three JEPs elevated from Candidate to Proposed to Target for JDK 27. The proposed release schedule has also been finalized. By Michael R

💬 Editor's Note

There was a flurry of activity in the OpenJDK ecosystem duri… （摘要兜底）

US's big bet on quantum computing may not be entirely legal

9/10 News

Deal also launched the first quantum foundry company, but is there a need for it?

Awarding Jay Haynes His Being Right Points for Predicting Apple Hitting $3 Trillion in Market Cap

9/10 News

Here’s a fun one. Back in 2014 I linked to a post by Jay Haynes in which he projected that with a very reasonable level of annual growth, Apple ought to reach a $3 trillion market cap within 10 years. At the time of his writing, Apple’s market cap was “just” $450 billion, and no company had hit the

面壁智能联合清华等开源中国首个基于华为昇腾训练的 1.58-bit 端侧大模型 BitCPM-CANN

9/10 New Product

Authenticity at scale used to be a contradiction. Not anymore. Define the brief. Set the style. Lum…

9/10 New Product

规模化的真实性曾是矛盾，如今已成现实。定义简报，设定风格，Luma Agents 从这里构建每一条 UGC 风格广告。让它真实 → http：//lumalabs.ai/app

Pope Leo calls for being 'profoundly human' in the age of AI

7/10 Industry

教皇里奥十四世在首份重要通谕《Magnifica Humanitas》中警告了AI及不受约束的技术力量带来的风险。该通谕于2026年5月15日发布，是教皇关于"在人工智能时代守护人"的宣言。文件重点讨论了AI驱动战争的危险、AI对劳动的影响，以及建立新的法律和伦理框架来规范技术的必要性，强调AI的快…

微软 Copilot Cowork 存在文件泄露问题

7/10 Tech

🕐 ~3 min read · Tech 7/10

不依赖新光刻工艺！麒麟 2027、昇腾 990 芯片在路上，华为何庭波"韬定律"论文要点大揭秘

💡 Detailed technical reference

华为何庭波在ISCAS 2026上提出"韬定律"，并介绍逻辑折叠（LogicFolding）技术。该技术通过三维空间拓扑重组提升芯片性能，不依赖新光刻工艺。在麒麟2026芯片测试中，晶体管密度从155 MTr/mm2提升至238 MTr/mm2，性能核心能效提高41%，最大时钟频率提升近13%。论文显示，麒麟2027芯片已进入Silicon状态，后续规划包括麒麟2028、2029。AI芯片方面，昇腾990计划在2030年左右引入逻辑折叠，硬件集成预计到2035年提高超过100倍。

🕐 ~3 min read · Tutorial 6/10

Dream-MPC: Gradient-Based Model Predictive Control with Latent Imagination

💡 Can be adapted into tutorial material

This paper introduces Dream-MPC, a hybrid reinforcement learning approach combining Model Predictive Control with learned models and policy priors. It addresses limitations of current methods by using gradient-based optimization for planning, effectively leveraging the advantages of both planning-based and policy-based paradigms to improve sample efficiency.

🕐 ~3 min read · Tutorial 6/10

Distilling Linearized Behavior into Non-Linear Fine-Tuning for Effective Task Arithmetic

💡 Can be adapted into tutorial material

This paper addresses expressivity limitations of linearized task vectors in model merging and unlearning. By distilling linearized behavior from non-linear models, it maintains compositionality benefits of linear fine-tuning while improving task vector expressiveness and reducing model interference.

🕐 ~3 min read · Tutorial 7/10

Harness， Scaffold， and the AI Agent Terms Worth Getting Right

💡 Can be adapted into tutorial material

本文旨在厘清 AI 智能体领域中易混淆的关键术语。文章指出，模型（如 Claude、GPT）本身是无记忆、无循环的大语言模型。其行为由"Scaffolding"（行为定义层，如系统提示、工具描述）塑造，而"Harness"（执行层）负责调用模型、处理工具调用与控制循环，是智能体运行的核心。两者结合，模型才能成为智能体。文章以 Claude Code、Codex 为例，说明同一模型搭配不同 Harness 会产生迥异体验，并提出了 Agent = Model + Harness 的常见理解框架。术语尚未统一，本文旨在提供一个实用的心智模型。

New Product

Grok Build is now available in Beta for all SuperGrok and X Premium+ users. Use Plan Mode， create i…

Grok Build 现已面向所有 SuperGrok 和 X Premium+ 用户推出 Beta 版。使用 Plan Mode，通过 Imagine 创建图像和视频，并使用 CLI 构建自动化或编排器。访问 http：//x.ai/cli 开始使用。

✅Implicit caching is now live on Qwen3.7-Max - kicks in automatically， no setup needed. ⚡️Faster + …

✅隐式缓存现已在Qwen3.7-Max上线--自动启用，无需设置。 ⚡️开箱即用，更快更便宜。需要更高、更确定的命中率？请尝试显式缓存。🙌 🔗最佳实践🔗：https：//www.alibabacloud.com/help/en/model-studio/explicit-cache-best-p…

Opinion

The Double Dilemma in Multi-Task Radiology Report Generation: A Gradient Dynamics Analysis and Solution

Analyzes the challenge in multi-task automatic radiology report generation of balancing discriminative clinical supervision constraints with generation smoothness requirements. Identifies failure modes in linear scalarization strategies and proposes gradient dynamics solutions for improved clinical AI systems.

Industry

Anthropic 新模型震动金融圈，欧洲央行紧急开会要求升级网络防御

OpenAI GPT-5.6 模型曝下月发布：AI 上下文 150 万 tokens

多名开发者在 OpenAI Codex 后端日志中发现未官宣的 GPT-5.6 模型，内部代号 iris-alpha。该模型将支持 150 万 token 的上下文窗口，较当前 GPT-5.5 的 105 万 token 提升约 43%，有望于今年 6 月发布。

Apple isn't just adding Gemini to Siri. It is reportedly using a custom 1.2T-parameter Google model …

据报道，苹果为改造下一代Siri，正使用一个定制版、参数规模达1.2T的Google大模型作为其核心，这显著大于预估约300B参数的Gemini 3.5 Flash。该模型将驱动Siri的部分功能，其中简单查询预期会在本地设备运行。苹果面临的关键挑战是确保该大模型能够足够快速地响应日常问题。

Tutorial

Anthropic co-founder Chris Olah's remarks on Pope Leo XIV's encyclical "Magnifica humanitas"

Anthropic联合创始人Chris Olah在梵蒂冈出席教皇Leo XIV关于AI的通谕发布会。他指出，所有前沿AI实验室都面临商业、研究及地缘政治等多重压力，这可能与做正确的事相冲突，因此外部监督至关重要。

self improvement prompt for codex

这是一个结构化的提示词，用于指导 Codex 自动分析其历史记录以识别并固化重复工作流。该框架要求 Codex 回顾会话、Memories 等数据，找出重复、耗时且有明确复用价值的任务。筛选标准包括至少出现两次、输入稳定、可提升效率等。

📭Skip Today

Auto-filtered. Here's why — so you know you're not missing out:

Agentic Proving for Program Verification
→ Single-source paper, low reader value
Memorization Dynamics of Fill-in-the-Middle Pretraining
→ Single-source paper, low reader value
Are Frontier LLMs Ready for Cybersecurity? Evidence for Vertical Foundation Models from Dual-Mode Vulnerability Benchmarks
→ Single-source paper, low reader value
HARNESS-LM: A Three-Phase Training Recipe for Harnessing SLMs in Sponsored Search Retrieval
→ Single-source paper, low reader value
Can AI Guess What You Know? Performance Comparison of Large Language Models for Human Domain Knowledge Estimation From Communication Logs
→ Single-source paper, low reader value
Same Model, Different Weakness: How Language and Modality Reshape the Jailbreak Attack Surface in Frontier MLLMs
→ Single-source paper, low reader value
Anthropic co-founder Chris Olah's remarks on Pope Leo XIV's encyclical "Magnifica humanitas"
→ Already covered, no new facts today
SciAtlas: A Large-Scale Knowledge Graph for Automated Scientific Research
→ Single-source paper, low reader value

Subscribe to Xiaohu AI Daily